This project is an simple example to demonstrate the usage of MPI for numerical computing. It implements a distributd alternating direction method of multipliers (ADMM) algorithm for solving lasso problem.

This repository provide implementations in three languages:

• C. The C implementation requires GNU Scientific Library (GSL) and Open MPI (or other MPI implementation).
• Python. The python version requires Numpy/SciPy and MPI4Py.
• Lua. The Lua version is implemented based on Torch7 with MPIT.

First of all, one should load the require modules on TACC before running the code. Using the follow script to load modules

. ./loadmod.sh

Note: the dot before "./loadmod.sh" is critical and cannot be eliminated.

Then, one can choose the following implementation to run.

C version

make

make run

Python version

make runpy

Lua version

make runlua

C version

make clean; make

sbatch my_c_lasso_job

Python version

sbatch my_py_lasso_job

Lua version

sbatch my_lua_lasso_job

Note: Here assumes Torch7 is installed in $HOME/torch/install.

Here is some very unofficial result on running time comparison of the three implementations on a Macbook Pro labtop computer and the TACC stampede node. The speed depends on the implementation of the MPI and numerial library.

The c implementation is from here.